Add PROMPT_SELECTION.md (prompt-selection methodology)

PROMPT_SELECTION.md

@@ -0,0 +1,48 @@
+# Prompt Selection Criteria
+
+`metadata.json` ships the final SAM 3 prompts produced by the two-stage pipeline
+described in the paper §3.1. This document records the criteria applied during
+stage 2 (Claude Sonnet review and refinement), to enable extension to new
+videos under the same protocol.
+
+## Goal
+
+For each video, select a single entity whose removal makes the most questions
+unanswerable, expressed as a short prompt (≤4 words) that SAM 3 can ground
+reliably.
+
+## Criteria applied during refinement
+
+1. **Evidence-bearing** — the entity must be the visual referent of at least
+   one question's correct answer; removing it should obstruct answering.
+2. **SAM 3-friendly** — short noun phrases with optional visual descriptor
+   (color or clothing). Examples: `boy in red`, `man with guitar`, `white dog`.
+3. **Singular and specific** — prefer specific descriptors over generic terms
+   (`man` over `person`); avoid plurals (SAM 3 cannot reliably target
+   "people" or "men" as a group).
+4. **Cross-question consensus** — if multiple questions reference the same
+   entity, that entity is a higher-confidence removal target.
+
+## Patterns excluded after audit
+
+| Issue | Resolution |
+|---|---|
+| Positional refs (`left`, `right`, `middle`) | Replace with visual descriptor or bare noun |
+| Relative refs (`taller`) | Drop to bare noun |
+| Role-based (`drummer`, `singer`) | Rewrite as entity (`man with guitar`) |
+| Object-as-target (`phone`, `bicycle`) | Switch to the person in the scene |
+| Generic `person` when a more specific term applies | Rewrite (`man`, `boy`, `girl`) |
+| Non-SAM-3 vocabulary (`polkadot`) | Simplify to bare noun (`baby`) |
+| Action-based descriptor (`standing with sticks`) | Rewrite to entity reference |
+
+## Stage 1 (rule-based) source
+
+The rule-based bootstrap selector (Stage 1) is implemented in
+`prompt_selector.py` in the companion code release; see paper §3.1. It parses
+each question's grammatical subject and scores candidate entities by
+subject-frequency, plural penalty, and presence of visual descriptors,
+producing an initial set of prompts that stages 2–3 refine against the
+criteria above. The released `metadata.json` reflects the post-verification
+prompts and is the canonical artifact — no need to re-run the selection
+pipeline to use the dataset; the script is only needed to extend the dataset
+to new videos.